You don't have an AI skills gap. You have a permission gap

Priya has been a data engineer at the same Melbourne fintech for four years. Good company, solid work, pipelines that don't fall over. Lately her phone keeps ringing with recruiters asking about AI and machine learning roles. And every time, she says a version of the same thing: I don't have a formal ML background. Is this even realistic?

Read that back. A recruiter, whose whole job is to find people who fit, looked at her and decided she was worth calling about an AI role. She is the one person in that conversation who doesn't believe it.

I'm the person who makes that call. Thirteen years on the ANZ hiring side, a lot of it recruiting top tech talent for Atlassian, has taught me to read the gap between what someone has built and what a role actually needs. When I ring a data engineer about an AI-adjacent role, I'm not being generous. I've already checked what the hiring manager needs against what this person has built, and judged the distance between the two closeable. The call is the verdict. Priya keeps treating it as a mistake.

That gap, between what the market already sees in you and what you'll let yourself claim, has a name. It isn't a skills gap. It's a permission gap.

The word "AI" is hiding four different jobs

Here's why the permission gap gets so wide, so fast. The advice everywhere is "learn AI or get left behind." It never tells you which AI, because keeping it vague is what keeps you buying the next course. So the word sits there, one giant fog, and you assume the scariest version of it is the one being asked of you.

It's really four jobs, and the distance between you and each one is completely different.

The first door is research and modelling. Building models from scratch, down to the maths underneath them. This is the lane that genuinely wants a PhD or a research record. It's real and it's hard, and almost never the role a recruiter is ringing a working engineer about. It's the door Priya pictures every time her phone lights up, and it's the one door actually shut to her today.

The second is ML engineering and MLOps. Taking models and making them run in production. Serving, monitoring, retraining, the pipelines around them. Look at what that actually is: deployment and infrastructure with some machine learning sitting on top. For a data engineer, that's most of what you already do, plus two things worth learning.

The third is applied AI, the fast-growing one. Building things on top of the models that already exist: LLM APIs, retrieval, agents, and the part that separates people, making it reliable and cost-controlled once real users are hitting it. There's no PhD here. The joke inside the field is that the entry requirement is an API key. What's scarce is the judgment to take a clever demo the last mile into something a business can trust. The maths is the easy part.

The fourth is ordinary engineering that now expects AI fluency. Your existing job, done by someone who's good with the tools. Eight in ten business leaders globally now say they're more likely to hire someone comfortable using AI tools than someone with more experience who isn't. This one is barely a move. It's the floor rising under everyone, and the engineers who act like it isn't are the ones I find hardest to defend on a shortlist.

Four doors. Priya has been standing in front of the only locked one, deciding the whole building is shut.

The honest gap is smaller than the fear

Say "AI" and the fear has nowhere to land, so it spreads into a wall. Name the actual door and it shrinks to a list you can finish.

For a data engineer moving toward MLOps or applied AI, the honest gap is small and boring. The fundamentals of how models get evaluated. Serving and monitoring what you deploy. For applied AI, add retrieval, agents, and the discipline of testing and costing a system that talks to a language model. For an experienced engineer that's a few months of focused work. People who already write good software have made this jump in a single quarter.

I want to be precise about what I'm not saying, because the market is full of people telling you the jump is free. It isn't. There's a real difference between a chatbot you built in a weekend and a system that stays reliable and affordable when a thousand people hit it at once. That difference is the actual gap. But you close it by building one real thing. A certificate won't.

Which brings us to the money, because Priya's other question is whether any of this moves her salary.

What actually moves the money

The number is real. PwC's 2026 AI Jobs Barometer puts the Australian wage premium for AI skills at 62 percent, up from 57 percent the year before, with AI-skilled workers earning a median around A$143,000 against A$104,000 for the wider workforce. Robert Half's 2026 data puts a Sydney AI engineer's median near A$185,000. AI Engineer is now the single fastest-growing job in the country.

Here's the part that trips people up. That premium does not attach to the letters A-I on your CV. I've interviewed people who relabelled themselves "AI Engineer" the week before applying, and the title falls apart the moment you ask them what they actually shipped and what it cost to run. A confident label with nothing under it gets tested hard, and it fails in the room, every time.

What clears the band is one deployed thing. A system with real users, a business reason to exist, and evidence you thought about whether it was reliable and what it cost. That's the evidence I'm paying for when I move someone from the data-engineer band into the applied-AI one.

So the honest answer to Priya was never about her headline. Build one real thing, and the number takes care of itself.

The test that tells you which gap you actually have

Try this before you sign up for anything. Name the last thing you put in front of a real person that had a model somewhere in the loop. A tool a teammate now uses. A workflow you wired up end to end and left running. A one-off prompt you typed into a chat window and forgot doesn't count.

If you can name one, stop. You're already through the door, and no course sells the thing you're actually missing, which is the nerve to say out loud what you've built.

If you can't name one, good, now you know the real gap, and it's the most closeable one in this whole piece. That's a weekend to start and a few months to get serious. No degree in it.

Either way, the move was the same. Go and build one real thing, then say plainly what it does.

The reason this is so hard to see on your own is that the permission gap is invisible from the inside. You can't tell whether you're genuinely under-qualified for a lane or just under-claiming it, because from where you're standing the two feel identical. That's the specific thing we built Career Direction in Careersy AI to do: take what you've actually done, name which of the four doors is already open to you, and lay out the short list standing between you and it, so you stop preparing for the one job nobody asked you to do. "Learn AI" is fog. "You're two skills off the applied-AI lane, and here they are" is something you can start tomorrow.

Priya's problem was never machine learning. It was permission, dressed up as a skills gap, because "I need to learn more" is an easier thing to sit with than "I'm ready and I'm scared to say so." The recruiters had already worked out which door was open for her. She was the last one to check the handle.

FAQ

Do I need a PhD or a formal ML background to work in AI?

Only for one lane. Research and modelling roles, building models from scratch, genuinely want that depth. The lanes most working engineers are actually recruited into, ML engineering and applied AI, do not. Applied AI in particular is built on top of existing models and rewards strong software skills and product judgment far more than a research background.

I'm a data engineer. Is moving into AI realistic, and what's the honest gap?

Yes, and it's one of the more natural moves in tech right now. Your pipeline and infrastructure skills map almost directly onto ML engineering and MLOps, the deployment-heavy side of AI. The honest gap is the machine-learning fundamentals and the model-serving and monitoring layer, months of focused work for an experienced engineer, not a full retraining.

Will adding AI skills actually increase my salary in Australia?

It can, meaningfully. PwC's 2026 Barometer put the AI-skills wage premium in Australia at 62 percent, and a Sydney AI engineer's median sits near A$185,000 on Robert Half's 2026 figures. But the premium attaches to demonstrated, deployed work, not to the word "AI" on your CV. A relabelled title with nothing shipped behind it gets found out in the interview.

Should I choose the IC track or move into leadership as I go into AI?

You don't have to decide yet, and deciding now is usually a mistake. The scarcest, best-paid thing in the market right now is hands-on proof you can ship reliable systems with a model in the loop. Bank six to twelve months of that evidence first, then let the lane choice be led by what you've built, not by a coin flip you make today.